Child Development, November/December 2015, Volume 86, Number 6, Pages 1948–1965

The Dimensionality of Language Ability in Young Children

Language and Reading Research Consortium

The purpose of this study was to empirically examine the dimensionality of language ability for young chil-
dren (4–8 years) from prekindergarten to third grade (n = 915), theorizing that measures of vocabulary and
grammar ability will represent a unitary trait across these ages, and to determine whether discourse skills rep-
resent an additional source of variance in language ability. Results demonstrated emergent dimensionality of
language across development with distinct factors of vocabulary, grammar, and discourse skills by third
grade, confirming that discourse skills are an important source of variance in children’s language ability and
represent an important additional dimension to be accounted for in studying growth in language skills over
the course of childhood.

Scholars within a number of disciplines (e.g.,
speech–hearing sciences, developmental psychol-
ogy, and linguistics) tend to view language ability
as a complex system consisting of multiple dimen-
sions. As a result, textbooks and discussion on lan-
guage development and disorders are often
organized by different dimensions of language (e.g.,
Bishop, 1997; Schwartz, 2009). The presumed multi-
dimensionality of language ability has practical
import as well. For instance, the diagnosis of disor-
This paper was prepared by a Task Force of the Language and
Reading Research Consortium (LARRC) consisting of Jill Penti-
monti (Convener), Ann O’Connell, Laura Justice, and Kate Cain.
LARRC project sites and investigators are as follows: Ohio State
University (Columbus, OH)—Laura M. Justice (Site PI), Richard
Lomax, Ann O’Connell, Jill Pentimonti, Stephen A. Petrill, and
Shayne B. Piasta; Arizona State University (Tempe, AZ)—Shelley
Gray (Site PI) and Maria Adelaida Restrepo; Lancaster University
(Lancaster, UK)—Kate Cain (Site PI); University of Kansas
(Lawrence, KS)—Hugh Catts (Site PI), Mindy Bridges, and Diane
Nielsen; University of Nebraska-Lincoln (Lincoln, NE)—Tiffany
Hogan (Site PI), Jim Bovaird, and J. Ron Nelson. Jill Pentimonti is
now at American Institutes for Research. Stephen A. Petrill was a
LARRC coinvestigator from 2010 to 2013. Hugh Catts is now at
Florida State University. Tiffany Hogan is now at MGH Institute
of Health Professions. J. Ron Nelson was a LARRC coinvestigator
from 2010 to 2012.
This work was supported by Grant R305F100002 of the Insti-
tute of Education Sciences’ Reading for Understanding Initiative.
We are grateful to the numerous staff members, research associ-
ates, school administrators, teachers, children, and families who
participated. Key personnel at study sites include: Garey Berry,
Beau Bevens, Jennifer Bostic, Shara Brinkley, Lori Chleborad,
Dawn Davis, Jaclyn Dynia, Michel Eltschinger, Tamarine Fore-
man, Rashaun Geter, Sara Gilliam, Miki Herman, Jaime Kubik,
Trudy Kuo, Gustavo Lujan, Carol Mesa, Denise Meyer, Maria
Moratto, Kimberly Murphy, Marcie Mutters, Trevor Rey, and
Stephanie Williams.
The views presented in this work do not represent those of the
federal government, nor do they endorse any products or find-
ings presented herein.
Correspondence concerning this work should be sent to Jill
Pentimonti, 1000 Thomas Jefferson St., Washington, DC 20007.
Electronic mail may be sent to jpentimonti@air.org.
dered language in children is based on evidence of
symptoms (delays or deficits) across one or more of
the dimensions of vocabulary, grammar, narrative,
discourse, or pragmatics (American Psychiatric
Association, 2012). Such practices further reinforce
the perspective that language ability comprises
multiple dimensions and that these dimensions rep-
resent independent abilities. Furthermore, it has
been shown that early weaknesses in vocabulary,
grammar, and discourse comprehension and pro-
duction are related to later reading comprehension
difficulties (Catts, Fey, Zhang, & Tomblin, 1999;
Nation, Cocksey, Taylor, & Bishop, 2010). Thus, our
understanding of the structure of language extends
beyond oral language to literacy and educational
attainment.
A premise relevant to the present study is that
such conceptualizations of language ability are la-
rgely theoretical in nature. To date, there has been
limited empirical assessment of the extent to which
the various theorized dimensions of language do
indeed represent latent abilities at a given point in
a child’s development. Importantly, there is limited
empirical justification for measuring language abil-
ity (and diagnosing language disorders) as a set of
distinct psychological dimensions (using subtests
for interpretation) or as a unitary psychological trait
(using the composite for interpretation; see Tomblin
© 2015 The Authors
Child Development published by Wiley Periodicals, Inc. on behalf of Society
for Research in Child Development.
This is an open access article under the terms of the Creative Commons
Attribution-NonCommercial License, which permits use, distribution and
reproduction in any medium, provided the original work is properly cited
and is not used for commercial purposes.
All rights reserved. 0009-3920/2015/8606-0020
DOI: 10.1111/cdev.12450

& Zhang, 2006). This study empirically examined
the dimensionality of language ability for consecu-
tive age groups from 4 to 8 years. Using a cross-
sectional sample and measures of the language
dimensions of vocabulary, grammar, and discourse,
study outcomes help to advance our fundamental
understanding of the nature of language ability
among children and inform practices that depend
on this understanding.
Dimensionality of Language Ability
Language is a complex system; thus, it comes as
no surprise that there are different ways to concep-
tualize its dimensions. One approach is to distin-
guish between language performance in different
modalities, that is, comprehension (what is under-
stood) versus production (how language is used).
In support of this approach, there are clear dissocia-
tions between comprehension and production dur-
ing early language development: For example,
children between 8 and 16 months understand
many more spoken words than they produce
(Bates, Dale, & Thal, 1995), and comprehension and
production result in specific and different patterns
of brain activity in 1-year-olds (Mills, Coffey-
Corina, & Neville, 1997). However, research
empirically examining the distinctiveness of the
comprehension versus production modality, at least
for children 5 years of age and older, fails to
support these modalities as true dimensions of
language ability (see Tomblin & Zhang, 2006).
Thus, the research reported here cuts across modal-
ity to examine the distinctiveness of theorized
dimensions of language, and considers whether
these should be conceptualized as a single system
or separate subsystems in the early primary grades.
The majority of research that speaks to this ques-
tion has examined the relation between vocabulary
and grammar. One school of thought is that vocab-
ulary and grammar represent separate modular
systems, each governed by different underlying
systems or mechanisms (Pinker, 1998). Certainly we
find that in languages such as English, one-word
production precedes the use of grammar in the
form of inflections on single words or combinations
of words. This temporal asynchrony in lexical and
grammatical development could be taken as evi-
dence that different modular learning mechanisms
for words and grammar exist. However, the com-
putational approach for simulating language devel-
opment has demonstrated that the same learning
mechanism could be used for both vocabulary and
grammar acquisition (Plunkett & Marchman, 1993).
Dimensionality of Language 1949
Furthermore, grammatical development depends on
a critical lexical base, and different phases of
grammar development are linked to different
phases of lexical growth (Marchman & Bates, 1994).
These two strands of evidence lead some to argue
that the apparent dimensionality of different lan-
guage systems might be an emergent, rather than
innate, property of a developing language system
(Elsabbagh & Karmiloff-Smith, 2006).
Empirical Investigations of Language Dimensionality
The period of language development examined
in this article is 4–8 years of age, which is typically
later than the age range examined in the modular-
ity debates on vocabulary and grammar (e.g., Bates
et al., 1995). During this period, extensive language
development occurs and many language and lan-
guage-related disorders are diagnosed, such as
specific language impairment (SLI) and dyslexia.
For these reasons, we consider it essential to under-
stand how best to conceptualize language within
this age range to inform not only theory, but also
practices such as assessment, intervention, and
instruction.
We know of only one prior empirical investiga-
tion of the dimensionality of language ability in
typically developing young children, conducted by
Tomblin and Zhang (2006). Using data collected as
part of a longitudinal epidemiological study of lan-
guage disorders among kindergarten children,
Tomblin and Zhang examined whether children’s
language ability at kindergarten, second, fourth,
and eighth grades represented a uni- or multidi-
mensional trait, and whether representation of this
trait (or traits) changed as children grew older. The
dimensions of language ability assessed were
vocabulary and grammar and were measured for
both modalities of comprehension and production.
From an initial sample of 1,929 kindergarteners,
about 600 children were followed over time
(Tomblin & Zhang, 2006). Modality (comprehension
and production) was best considered to be unidi-
mensional. Critically, there was evidence of
emergent dimensionality: In the linguistic dimen-
sion, the correlations between factors representing
vocabulary and grammar at each grade (kinder-
garten, second, fourth, and eighth grades, respec-
tively) showed that these were high (e.g., rs = .94,
.93, .90, and .78 with increasing grade). However,
the lower value (.78) seen at eighth grade sug-
gested that language ability might be considered
multidimensional in the later grades. The possibil-
ity of emergent dimensionality was supported by
1950 LARRC
confirmatory factor analysis. At the kindergarten,
second grade, and fourth grade time points, the
one-dimensional model fit well and represented the
most parsimonious approach to conceptualizing
dimensionality of language ability at these grades.
In contrast, at eighth grade, the two-dimensional
model separating vocabulary and grammar was the
better fitting model. This work suggests that a uni-
dimensional model may not be the most accurate
description of language abilities across develop-
ment.
Language Dimensionality: A Broader Perspective
Tomblin and Zhang (2006) note that they did not
include measures of language that are often
included in the dimension of discourse (or text)
and, therefore, that their findings cannot be general-
ized more broadly. An understanding of the struc-
ture of language including discourse is critical for
both theoretical and practical reasons. Theoretically,
vocabulary and grammar have been conceptualized
as lower level skills that are important, but not suf-
ficient, for discourse comprehension, which has
been called a higher level dimension of language
(Kim, 2014; Lepola, Lynch, Laakkonen, Silven, &
Niemi, 2012). Measures of discourse include the
ability to understand narrative, generate inferences,
and monitor comprehension language. Notably,
these discourse-level language skills develop early:
For example, comprehension monitoring during a
narrative is evident between 30 and 39 months
(Skarakis-Doyle & Dempsey, 2008) and infants’
early word learning depends on inferential process-
ing (Tomasello, Carpenter, & Liszkowski, 2007).
Furthermore, when discourse-level language skills
such as inference, narrative skills, and comprehen-
sion monitoring are included alongside vocabulary
and grammar, each predicts variance in listening
comprehension in 4- to 6-year-olds (Kim, 2014;
Lepola et al., 2012; Silva & Cain, 2014). This litera-
ture suggests that if the dimension of discourse
were included in an examination of language
dimensionality, it might form a separable dimen-
sion to vocabulary and grammar. On the practical
side, there is a strong body of work demonstrating
the language bases of literacy outcomes (Berninger
& Abbott, 2010; Catts, Hogan, & Fey, 2003; Hogan,
Cain, & Bridges, 2012). More accurate models of
language structure in prereaders and beginner read-
ers therefore have broader implications beyond sim-
ply language development and difficulties: It might
better inform our identification of children at risk of
later literacy difficulties.
Aims of the Present Study
The overarching aim of the present study was
to further our understanding of the dimensional-
ity of language ability among children in the 4-
to 8-year-old age range. Two interests were
addressed, the first of which was to examine
dimensionality of language ability with respect to
vocabulary and grammar. Prior work has indi-
cated that these two dimensions of language rep-
resent a unitary trait across these ages (see
Tomblin & Zhang, 2006); however, a distinguish-
ing feature of the present work is that we extend
the study of language dimensionality to children
in prekindergarten and include a wider collection
of multiple indicators of language ability for each
proposed dimension to derive latent variables.
The second interest was to determine if discourse
skills represent an additional source of variance
in language ability that is unique from vocabulary
and grammar at these ages. Despite theoretical
representation in the literature of the distinctive-
ness of vocabulary and grammar versus discourse
language skills in the literature on reading devel-
opment (Cain, Oakhill, & Bryant, 2004; Dickinson,
Golinkoff, & Hirsh-Pasek, 2010; Kim, 2014; Lepola
et al., 2012), the present work is the first effort of
which we are aware to empirically determine
whether these theorized dimensions in fact repre-
sent valid dimensions of language ability. In addi-
tion, we assessed whether these dimensions
become more salient across development, as
found by Tomblin and Zhang (2006), theorizing
that the emergence of discourse language skills as
a dimension of language ability might only be
observed for the older children in our sample.
Method
Data for this study were collected as part of a mul-
tisite, 5-year longitudinal research project conducted
by the Language and Reading Research Consortium
(LARRC). The purpose of the LARRC longitudinal
study was to identify and model language pro-
cesses important for reading comprehension in chil-
dren enrolled in prekindergarten (PK) to Grade 3
(G3). The longitudinal study featured a cohort
design, such that in Year 1 five cohorts of children
in each of PK and G3 were ascertained and then
followed until G3; children thus graduated from
the study annually. The study design thus yields
both cross-sectional data and longitudinal data. In
this study, we utilized the cross-sectional data col-

lected for five waves of children (PK–G3) during
the 2011–2012 academic year.
Participants
Participants were 915 children in PK through G3
recruited from 69 schools and/or preschool centers
at four data collection sites (Arizona, Kansas,
Nebraska, Ohio). To recruit participants into the
study, teachers provided children’s caregivers with
recruitment packets that contained literature on
LARRC, a caregiver consent form, a family screener
questionnaire, and a return envelope. Caregivers
could return the forms directly back to the teacher
or mail them to the research team. This process was
used until each site recruited up to 110 PK children
and at least 30 (but no more than 50) children from
each of K–G3; the number of participants per site
was designed to be approximately equal. The chil-
dren in our sample were similar to the relative pop-
ulation of children at recruitment sites across key
demographic characteristics (e.g., eligibility to
receive free/reduced price lunch and membership
in racial/ethnic categories). Upon receipt of consent
for a given child, research staff members reviewed
the family screener to ensure that consented chil-
dren could appropriately participate in assessments
selected for the general sample. This included being
rated as understanding and speaking English flu-
ently and not having severe or profound disabili-
ties. In addition, for PK children to be enrolled in
the sample, it was expected that they would matric-
ulate to K the following year. In instances when
consent was received for the sibling of a selected
child, they were excluded from the sample.
Of the 915 participating children, 420 were in PK,
124 were in kindergarten (K), 125 were in Grade 1
(G1), 123 were in Grade 2 (G2), and 123 were in G3.
The PK cohort was purposefully designed to be the
largest, as it would be followed up for 5 years and
attrition was expected. Children’s ages ranged from
4 years 6 months to 8 years 6 months. The sample
included slightly more boys than girls (53% vs. 47%)
and the majority of children were White/Caucasian.
Specifically, 85% of children were White/Caucasian,
8% were Black, 5% were Asian, and 2% were Other.
About 10% of children were Hispanic. Seventy-eight
percent of families reported speaking primarily
English at home; other languages spoken at home
included Spanish, Chinese, Amharic, and Viet-
namese. Seventy percent of children resided in
two-parent households. Nearly 10% of children had
individual education plans, and about 16% qualified
for free/reduced lunch.
Dimensionality of Language 1951
Procedure
The primary procedure for this study, beyond
recruiting the longitudinal cohorts of children,
was that of assessment. Children completed a
comprehensive assessment battery in the latter
half of the academic year (January through May).
The battery required an average of 5–6 hr to
complete, with measures administered in 15- to
40-min blocks. For the measures utilized in the
present study, all children were assessed individu-
ally by trained assessors in quiet locations at their
schools, a university laboratory, or other public
facility (e.g., library). Assessors were certified for
a given measure following completion of an
extensive standardized training program that
included, for instance, completion of a written
quiz concerning administration and scoring proce-
dures (required 100% correct), and completion of
two live administrations that were observed by
an experienced assessor (90% accuracy to adminis-
tration and scoring procedures based on rubrics
developed for this purpose). Measures that were
deemed to be too complex to be scored reliably
in the field were audio recorded using digital
recorders and scored by trained research staff
members subsequently in a laboratory setting.
Measures
The present study utilized a subset of measures
from the larger assessment battery that represent
children’s language ability with regard to vocabu-
lary, grammar, and discourse skills. Table 1 pro-
vides a list of measures administered by grade, as
well as information on which measures were audio
recorded and postscored for scoring accuracy. Inter-
rater reliability was acceptable for all postscored
measures, with intraclass correlations ranging from
0.86 to 0.99 across measures and grades. Missing
data percentages ranged from 0% to 30%. The aver-
age missing data percentage across all grades and
all measures was 7.96%.
Vocabulary
Three standardized measures of vocabulary were
administered. Among the vocabulary measures uti-
lized across all grades, all sample-specific reliabili-
ties exceeded .70 with the exception of the Clinical
Evaluation of Language Fundamentals (CELF)
Word Classes subtest for the G3 sample (with
a = .69). Descriptive statistics are provided in
Table 1.
1952 LARRC

Table 1
Descriptive Statistics (M, SD) by Measure and Grade

PK K G1 G2 G3

Lower level language (LLL)/Grammar (G)

CELF–IV—Word Structure (CFWS)a 15.59 (5.56) 20.49 (5.22) 24.39 (3.96) 26.23 (3.58) 27.46 (2.90)
CELF–IV—Recalling Sentences (CFRS) 32.11 (13.9) 42.63 (13.94) 55.67 (13.62) 60.47 (13.72) 65.96 (13.86)
TEGI—Past tense probe (TEGT) 8.52 (4.31) 9.31 (4.84) — — —
TEGI—Third-person singular probe (TEGS) 6.99 (2.93) 7.56 (2.75) — — —
TROG (TRG) 6.24 (3.76) 9.76 (4.27) 13.86 (3.34) 14.41 (3.58) 15.54 (2.90)
Morphological derivation — — 7.89 (4.21) 10.57 (4.56) 13.72 (4.04)
LLL/Vocabulary (V)
PPVT 93.78 (19.31) 111.46 (18.59) 129.15 (16.95) 137.80 (16.68) 150.86 (16.81)
EVT 70.06 (13.79) 82.57 (12.11) 96.94 (14.01) 105.75 (13.77) 113.84 (14.28)
CELF–IV—Word Classes Receptive (CFWCr)a 14.29 (4.44) 17.55 (2.91) 18.97 (1.93) 19.77 (1.26) 11.31 (3.21)a
CELF–IV—Word Classes Expressive (CFWCe)a 7.82 (4.54) 12.61 (4.16) 14.90 (2.70) 16.38 (2.48) 6.44 (2.72)
Discourse (D)
Knowledge Violations Task 4.86 (2.88) 5.67 (2.86) — — —
Detecting Inconsistencies Task — — 4.45 (2.18) 4.75 (2.22) 5.50 (1.74)
Picture Arrangement Task 2.61 (2.90) 5.90 (3.75) 8.95 (2.64) — —
Sentence Arrangement Task — — — 0.87 (1.17) 1.12 (1.26)
Inference Background Knowledge (InfBK)a 0.76 (0.40) 1.13 (0.40) 1.16 (0.42) 1.20 (0.44) 1.68 (0.29)
Inference Integration (InfInt)a 0.87 (0.49) 0.97 (0.45) 1.07 (0.43) 1.13 (0.48) 1.39 (0.39)

Note. Word Classes 1 was administered to children in PK to G2 (scores ranged from 0 to 21 for the receptive subtests and from 0 to 18 for
the expressive subtest) and Word Classes 2 was administered to children in G3 (scores ranged from 0 to 19 for the receptive subtests and
from 0 to 13 for the expressive subtest). Standard scores from the PPVT and EVT are available from the first author. EVT = Expressive
Vocabulary Test; PK = prekindergarten; K = kindergarten; G1 = Grade 1; G2 = Grade 2; G3 = Grade 3; CELF = Clinical Evaluation of
Language Fundamentals; PPVT = Peabody Picture Vocabulary Test; TEGI = Test of Early Grammatical Impairment; TEGS = third-person
singular probe; TRG = Test for Reception of Grammar; TROG = Test for Reception of Grammar–Version 2. aPostscored measure.

Peabody Picture Vocabulary Test. The Peabody
Picture Vocabulary Test, 4th ed. (Dunn & Dunn,
2007), specifically Form A, assessed children’s com-
prehension of single words. The assessor asked the
child to point to the picture, of four, that matched a
verbally presented stimulus word. The total raw
score was utilized in analyses.
Expressive Vocabulary Test. The Expressive
Vocabulary Test, 2nd ed. (Williams, 2007), Form A,
assessed children’s productive vocabularies. The
assessor asked the child to provide a single word
label for a picture or to provide a single word syn-
onym for a target word. The total raw score was
utilized in analyses.
Word Classes subtest (CFWCr and CFWCe). The
Word Classes subtests of the CELF, 4th ed. (Semel,
Wiig, & Secord, 2003) assessed children’s abilities to
understand relationships among words related to
semantic class features and produce these similari-
ties and differences verbally. Word Classes 1 was
administered to children in PK through G2, and
Word Classes 2 was administered to children in G3.
To administer this task, the assessor read a set of
words, which were also represented pictorially for
the younger children assessed with Word Classes 1.
To assess comprehension of words, the child was
asked to select the two words that went together.
For the productive portion, the child was asked to
verbalize how the two words were related. The
total number of correct responses on both portions
was tallied to compute the raw score utilized in
analyses.
Grammar
Six standardized measures of grammar were
administered. Different measures were appropriate
for different age ranges; thus, not all grade cohorts
received the same measures, as shown in Table 1.
Among the grammar measures utilized across all
grades, all sample-specific reliabilities exceeded .70
with the exception of the CELF Word Structure
subtest for the G3 sample (with a = .63).
Word Structure subtest (CFWS). The Word Struc-
ture subtest of the CELF (Semel et al., 2003)
assessed children’s abilities to apply word structure
rules to indicate inflections, derivations, and com-
parison and to select appropriate pronouns to refer
to people, objects, and possessive relationships. The
child listened to a model sentence and then

supplied a missing word, using appropriate inflec-
tional morphology, in a second sentence that mir-
rored the first. All children completed this subtest,
although a stop rule of eight incorrect responses
was utilized for PK children. The total number of
correct responses was tallied as the raw score.
Recalling Sentences subtest (CFRS). The Recalling
Sentences subtest of the CELF (Semel et al., 2003)
assessed children’s abilities to listen to spoken sen-
tences of increasing length and complexity and
repeat these sentences without changing word
meanings, inflections, derivations or comparisons,
or sentence structure. All children completed this
subtest, although two items were added to better
accommodate PK children. Specifically, the first two
items from the Recalling Sentences subtest of the
CELF: Preschool, II edition (Wiig, Secord, & Semel,
2004) were administered as the first two test items,
followed by the test items on the Recalling Sen-
tences subtest of the CELF in the designated order.
For each item, the child was asked to listen to a
sentence read by the assessor and to repeat it ver-
batim. Total points were summed to create the raw
scores utilized in analyses.
Past tense probe (TEGT). The past tense probe of
the Rice/Wexler Test of Early Grammatical Impair-
ment (TEGI; Rice & Wexler, 2001) assessed chil-
dren’s production of regular and irregular past
tense verbs, and was administered to PK and K
children only. The child was presented with a pic-
ture and verbal description of a boy or girl com-
pleting an action using the present participle verb
form. The child was then asked to respond as to
what the boy or girl did, requiring use of the past
tense form of the verb. The total raw score was uti-
lized in analyses.
Third-person singular probe. The third-person sin-
gular probe of the TEGI (Rice & Wexler, 2001)
assessed children’s abilities to produce /-s/ or /-z/
in present tense verb forms with singular subjects.
This probe was administered to PK and K children
only. The child was presented with a picture and
name of a person with a specific job (e.g., teacher)
and asked what that person does (e.g., teaches).
The total raw score was utilized in analyses.
Test for Reception of Grammar. The Test for
Reception of Grammar–Version 2 (Bishop, 2003)
assessed children’s comprehension of English gram-
matical contrasts marked by inflections, function
words, and word order. Test items are arranged in
20 blocks of four, each block assessing knowledge
of the same grammatical contrast. For each item,
the child was shown four pictures and the assessor
read the accompanying sentence and asked the
Dimensionality of Language 1953
child to point to the picture for that sentence. For
each sentence, one picture represented the target
meaning and the other three pictures represented
grammatical or lexical foils. The child was awarded
1 point if all item responses within a block were
correct; the total number of correct blocks was used
for analyses.
Morphological Derivation Task. A Morphological
Derivation Task described by Wagner (n.d.)
assessed children’s knowledge of derivational mor-
phology. Only children in G1–3 completed this
measure. The assessor presented the child with a
base word (e.g., farm) and an incomplete sentence
for which he or she provided a derived form of the
base word (e.g., My uncle is a ______). Responses
were tallied to provide the raw score utilized in
analyses.
Discourse
Five measures of discourse skills were adminis-
tered to examine comprehension monitoring, text
structure knowledge (specific to narratives), and
inference. Measures given differed slightly by
grade, as shown in Table 1. We provide reliability
data for these researcher-designed measures.
Comprehension monitoring—Knowledge Violations
Task. A researcher-developed measure based on
work of Cain and Oakhill (Cain & Oakhill, 2006;
Oakhill & Cain, 2012) was used to assess compre-
hension monitoring in PK and K children. The child
listened to short stories that were either entirely
consistent or included inconsistent information, as
in a story about a rabbit in a vegetable patch who
“especially liked the ice cream that grew there.”
After each story, the child was asked whether the
story made sense and, if not, what was wrong with
the story. The task included five practice stories
and seven test stories, five of which included incon-
sistent information. Children received 1 point for
each inconsistent story for which they correctly
answered the sense question and correctly identi-
fied the incorrect information in the story. These
points were tallied to create the raw score utilized
in analyses (possible range = 0–10). Internal consis-
tency measures (Cronbach’s alphas) in the present
sample ranged from .80 to .81 across grades.
Comprehension monitoring—Detecting Inconsisten-
cies Task (DI). A researcher-developed measure
based on the work of Cain and Oakhill (Cain &
Oakhill, 2006; Oakhill & Cain, 2012) was used to
assess comprehension monitoring for children in
G1–G3. The tasks included 5 practice stories and 12
test stories that were either entirely consistent or
1954 LARRC
included inconsistent information; 8 test stories
were inconsistent. Similar to the Knowledge Viola-
tions Task (KVT), the child listened to short stories
and was asked whether the story makes sense and,
if not, what was wrong with the story. The task
was more difficult than the KVT, because identifica-
tion of inconsistent information required integration
of two elements across two separate sentences (e.g.,
On her way she lost her purse. When she got to the
store, she took out her purse and bought her favorite
candy). Children received a point for each inconsis-
tent story for which they correctly identified the
incorrect information within the story. These points
were tallied to create the raw score utilized in anal-
yses (possible range = –8). Internal consistency
measures (Cronbach’s alphas) in the present sample
were .70 for G1, .73 for G2, and .54 for G3.
Narrative structure—Picture Arrangement Task. The
first 13 items of the Picture Arrangement Test from
the Wechsler Intelligence Scale for Children, 3rd ed.
(Wechsler & Kort, 2005) were adapted to create a
measure that assessed children’s text structure
knowledge, specific to ordering narrative events
into a causally and temporally coherent sequence.
This task was administered to children in PK
through G2. The child was told that she would
view some pictures that tell a story, but the story is
out of order. The assessor then showed the child a
set of three to five picture cards in a fixed order
and read a sentence that described each. The child
was asked to rearrange the pictures to put them in
the correct sequence. One practice item and 12 test
items were administered; a ceiling rule of five incor-
rect stories was applied. Internal consistency mea-
sures (Cronbach’s alphas) in the present sample
ranged from .76 to .88 across grades.
Narrative structure—Sentence Arrangement Task. A
researcher-developed measure based on work by
Oakhill and Cain (2012) and Stein and Glenn
(1982) assessed older children’s text structure
knowledge, specific to ordering narrative events
into a causally and temporally coherent sequence.
This measure was administered to children in G2
and G3. The child was told that she would read
some sentences that tell a story, but the story is
out of order. The assessor then showed the child a
set of 6–12 cards, with one sentence typed on each
card, in a fixed order and read each sentence
aloud to the child. The child was asked to rear-
range the sentences to put them in the correct
sequence. This measure consists of one practice
story and four test stories. Internal consistency
measures (Cronbach’s alphas) in the present sam-
ple were .67–.69.
Inference (InfBK and InfInt). A researcher-devel-
oped measure based on work by Cain and Oakhill
(2006) and Oakhill and Cain (2012) was used to
assess children’s ability to generate two types of
inferences from short narrative texts: inferences that
require integration of two premises, and inferences
that require integration of information in the text
with background knowledge to fill in missing details.
Following administration of a practice story, chil-
dren listened to two stories, after which the assessor
asked eight questions, reflecting four questions per
inference type. The total number of correct responses
for the two types of inference were summed and
averaged to produce two scores, one for background
knowledge and one for integration (possible
range = 0–2 for each). Internal consistency measures
(Cronbach’s alphas) for the averaged score across all
items in the present sample ranged from .64 to .78
across grades.
Results
Means and standard deviations for the primary
study measures are shown in Table 1. As expected,
mean scores tended to increase across grades.
Through inspection of skewness and kurtosis crite-
ria, histograms, and boxplots of the data, the major-
ity of variables showed adequate distribution with
no severe departures from normality. Furthermore,
no extreme outliers were identified within the data
across any of the grades.
Two interests were addressed in the primary
analyses. The first interest was to examine dimen-
sionality of language ability with respect to vocabu-
lary, grammar, and discourse skills as a unitary
construct or as two distinct dimensions consisting
of lower level language (LLL) (vocabulary and
grammar) skills versus discourse skills. The second
interest was to examine if discourse, vocabulary,
and grammar skills represent three distinct dimen-
sions of language ability. Measures indicating each
of the three language dimensions assessed are iden-
tified in Table 1. Both interests were addressed
simultaneously by examining three models to deter-
mine the best conceptualization of language dimen-
sionality across the grades PK to G3. Specifically,
we fitted a taxonomy of latent variable models
allowing for competing configurations of dimen-
sionality. These confirmatory factor analyses (CFA)
were conducted in MPlus v7.11 (Muthen & Muthen,
1993–2012) using the clustering option and maxi-
mum likelihood estimation with robust standard
errors to adjust for nonindependence within class-

rooms and slight non-normality of the data. Three
models were compared at each grade for quality of
fit: a unidimensional model of language, a two-
dimensional model differentiating a LLL factor
(vocabulary and grammar) from a factor represent-
ing discourse, and a three-dimensional model sepa-
rating vocabulary, grammar, and discourse ability.
Each lower dimension model can be obtained from
the higher dimension models by fixing the corre-
sponding factor correlations to 1.0; as such, these
three models form a set of nested models and rela-
tive model fit can be assessed through the chi-
square difference test. In each confirmatory model
we allowed for methods effects for measures from
the same instrument (e.g., both the comprehension
and production parts of the Word Classes, and also
for the One-Word Picture Vocabulary Tests), but no
other modifications were made. A graphical depic-
tion of these confirmatory models for the PK data
can be found in Figure 1; similar models were con-
structed for corresponding measures in K and each
of G1 through G3.
Recent recommendations on model fit and
assessment of model quality have called for a con-
sideration of several fit criteria rather than a strict
reliance on any single criteria, and this is the strat-
egy through which we evaluated our models (e.g.,
Hancock & Mueller, 2010, 2013; Hu & Bentler, 1995,
1999; Kline, 2013; Lomax, 2013; McCoach, Black, &
O’Connell, 2007; Moran, Marsh, & Nagengast,
2013; Tomarken & Waller, 2003, 2005). Correspond-
ingly, we used several approaches in our assess-
Figure 1. Graphical depiction of confirmatory models for the prekindergarten data. L = language; LLL = lower level language;
D = discourse; G = grammar; V = vocabulary. The figure is an exemplar of confirmatory models, as some measures change across
grade (see Table 1 for list of measures).
Dimensionality of Language 1955
ment of language dimensionality for each grade.
First, we considered a combination of absolute, par-
simony correction, and comparative indices of
model fit (Brown, 2006; Byrne, 2012). These
included the root mean square error of approxima-
tion (RMSEA), which ranges from 0 to 1; values
< 0.08 or, more conservatively, 0.05, suggest accept-
able model fit (Browne & Cudeck, 1993; MacCallum
& Austin, 2000). We report 90% confidence inter-
vals for the RMSEA and results of the closeness of
fit test (Browne & Cudeck, 1993), which tests the
null hypothesis that RMSEA ≤ 0.05; this test should
be ns. We also used the comparative fit index (CFI),
an incremental measure that compares the fit of the
theoretical model to a null model which assumes
the indicator variables in the model are uncorre-
lated (i.e., an independence model). CFI values
> 0.90 indicate a good fitting model (Hu & Bentler,
1999; Lomax, 2013; Schumacker & Lomax, 2004).
Comparison of nested models based on the CFI can
provide reasonable support for the more parsimo-
nious model if the incremental change in CFI < 0.01
(Moran et al., 2013). The standardized root mean
square residual (SRMR) is an absolute measure of
model fit and is the standardized difference
between the observed and predicted correlations; a
model that fits well to the data should have an
SRMR < 0.05 (Byrne, 2012). Model fit was also eval-
uated using Akaike’s information criteria (AIC), a
log-likelihood measure of fit useful for comparing
competing (typically non-nested) models and for
which smaller AICs indicate better fit (Kline, 2013).
1956 LARRC
In addition to examination of the model fit
indices, we used the Satorra–Bentler rescaled chi-
square difference test for statistically comparing
nested models (S–B v2; Satorra & Bentler, 1994). If
the scaled chi-square difference is not statistically
significant, the more parsimonious (i.e., the nested
or more restrictive) model is retained as having
model fit no worse than the more complex model.
If the difference is statistically significant, we are
not able to retain the more parsimonious model.
In evaluating model quality, we note that
descriptive fit statistics as well as results of statisti-
cal comparisons of nested models are susceptible to
sample size and other design features (McCoach
et al., 2007; Tomarken & Waller, 2005). Thus, as
advocated in the literature (e.g., Lomax, 2013; Muel-
ler & Hancock, 2010), we viewed the collection of
model fit indices as a guide rather than rely on a
single indicator or test as indicative of the “best”
representation of language. We also considered sub-
stantive interpretation of the CFAs, particularly
regarding discrimination between latent factors for
models with increasing dimensionality.
Accordingly, in our final approach to examining
support for increasing dimensionality across grades,
we considered the extent to which the latent factors
in the multidimensional models represented distinct
components of language. For each competing
model by grade, we reviewed the factor correla-
tions. We also calculated the amount of variance in
a measure that is explained by a factor by squaring
the standardized factor loadings and then averaged
these to estimate the average variance extracted
(AVE; Hair, Black, Babin, & Anderson, 2010; Nete-
meyer, Bearden, & Sharma, 2003). Comparing the
degree to which measure variance is captured by a
factor (AVE) relative to the shared variance
between constructs (squared factor correlations
[SFCs]) for the two- and three-dimensional models
provides information about how well the constructs
are distinct within each model. If the proportion of
variance extracted by a factor is less than the pro-
portion of variance shared between those factors,
there is little support for discrimination between
the factors (Fornell & Larcker, 1981; Hair et al.,
2010; Netemeyer et al., 2003). Consequently, our
decisions on the best dimensional representation of
language for each grade were based on the overall
constellation of model fit criteria, consideration of
discriminability between latent factors, and tests of
nested model comparisons.
Model fit statistics and results for the adjusted
chi-square comparison tests across models within
each grade are provided in Table 2. Table 3
contains information on the AVE for the set of
measures theoretically presumed to relate to each
factor in the models, as well as the SFCs for the
two- and three-dimensional models. In what fol-
lows, we present results for each of the five
grades examined in this study. We provide the
greatest depth of interpretation for the PK results,
as details provided therein can then be extrapo-
lated for the other grades. Construct reliability
(coefficient H) for the final model we selected
within each grade is also reported (Hancock &
Mueller, 2001).
Prekindergarten: Dimensionality of Language Ability
Model Fit Criteria
The constellation of fit statistics (see Table 2) was
very similar across the three models of dimension-
ality being compared. For example, RMSEA was
lowest for the two-dimensional model at 0.062, but
the confidence intervals for all three models were
similar, and the RMSEA estimates for the unidi-
mensional and three-dimensional models were simi-
lar at 0.065 and 0.063, respectively. Corresponding
tests for closeness of fit (i.e., RMSEA ≤ 0.05) were
not significant for any of the three models, with p-
close values of .019, .045, and .036, respectively.
Nonstatistically significant values for the “close-
ness” test indicate acceptable fit, although Brown
(2006) mentions that there is a tendency among
methodologists to prefer a p-close value > .50 in
order to claim that the RMSEA is sufficiently small.
None of these models achieved that boundary.
Among the three models for PK, the CFI and SRMR
values were the same. The work of Moran et al.
(2013) suggests that the more parsimonious of these
models can reasonably be supported given the lack
of incremental change in the CFI. Overall, however,
this collection of fit indices did not point to
unequivocal support for one model of language
dimensionality at PK over another.
Nested Model Comparisons
Next, we examined the (adjusted) S–B v2
difference tests for each pair of models (see
right-hand section of Table 2, column titled
Adjusted Dv2). This test assesses whether the fit
of the simpler (lower dimensioned) model is
appreciably worse than that of the more complex
(higher dimensioned) model. Results indicated
that the one-dimensional model provided slightly
worse fit relative to the two-dimensional and
 Dimensionality of Language 1957

Table 2
Results of Confirmatory Models by Grade

MLR scaling RMSEA, p close Adjusted r for latent

Model v2 df p factor (90% CI) CFI SRMR AIC Δv2 variables

Prekindergarten (n = 420)
Unidimensional 170.84 62 < .001 1.02 0.065, p = .019 [0.053, 0.076] 0.96 0.04 2,6387.03 —
Two-dimensional 159.70 61 < .001 1.02 0.062, p = .045 [0.050, 0.074] 0.96 0.04 2,6376.68 9.13a,** rLLL-D = .91
Three-dimensional 157.40 59 < .001 1.02 0.063, p = .036 [0.051, 0.075] 0.96 0.04 2,6379.11 13.34b,** rG-D = .90
1.88c rV-D = .94
rG-V = .99
Kindergarten (n = 124)
Unidimensional 60.53 62 .529 1.01 0.000, p = .939 [0.000, 0.052] 1.00 0.04 7,727.97 —
Two-dimensional 59.73 61 .522 1.00 0.000, p = .935 [0.000, 0.053] 1.00 0.04 7,728.24 0.89a rLLL-D = .95*
Three-dimensional 56.58 59 .565 1.00 0.000, p = .944 [0.000, 0.051] 1.00 0.04 7,729.30 3.75b rG-D = .92*
3.29c rV-D = .98*
rG-V = .98*
G1 (n = 125)
Unidimensional 85.25 52 .002 0.99 0.072, p = .101 [0.043, 0.098] 0.96 0.05 6,844.67 —
Two-dimensional 71.86 51 .029 1.01 0.057, p = .332 [0.018, 0.086] 0.98 0.05 6,834.08 17.82a,** rLLL-D = .85*
Three-dimensional 68.03 49 .037 1.00 0.056, p = .362 [0.014, 0.086] 0.98 0.05 6,833.88 18.43b,** rG-D = .82*
3.83c rV-D = .95*
rG-V = 1.06*,d
G2 (n = 123)
Unidimensional 87.66 52 .001 0.90 0.080, p = .073 [0.046, 0.101] 0.95 0.06 6,489.01 —
Two-dimensional 68.75 51 .049 0.90 0.053, p = .412 [0.003, 0.083] 0.98 0.05 6,473.85 17.97a,** rLLL-D = .80*
Three-dimensional 67.21 49 .043 0.90 0.055, p = .379 [0.010, 0.085] 0.98 0.05 6,476.52 20.45b,** rG-D = .79*
1.51c rV-D = .86*
rG-V = 1.03*,d
G3 (n = 122)
Unidimensional 87.74 52 .001 1.02 0.075, p = .071 [0.047, 0.102] 0.94 0.06 6,484.43 —
Two-dimensional 79.33 51 .010 1.02 0.067, p = .158 [0.036, 0.095] 0.96 0.06 6,477.61 7.51a,** rLLL-D = .78*
Three-dimensional 71.33 49 .020 0.99 0.061, p = .265 [0.025, 0.090] 0.97 0.05 6,471.82 12.87b,** rG-D = .80*
6.19c,* rV-D = .70*
rG-V = .90*

Note. MLR = robust maximum likelihood; RMSEA = root mean square error of approximation; CFI = comparative fit index;
SRMR = standardized root mean square residual; AIC = Akaike’s information criteria; LLL = lower level language.
a
Unidimensional versus two-dimensional models (df = 1). bUnidimensional versus three-dimensional models (df = 3). cTwo-dimensional
versus three-dimensional models (df = 2). dInadmissible parameter estimate.
*p < .05. **p < .01.

three-dimensional models, v2(1) = 9.13, p < .01
and v2(3) = 13.34, p < .01, respectively, and there
was no statistical difference between the two-
and three-dimensional models, v2(2) = 1.88,
p > .05. Thus, statistically speaking, the two-
dimensional model would be supported. Given
the known problems with distribution and sensi-
tivity to sample size of the chi-square difference
test (Tomarken & Waller, 2003), we next exam-
ined the factor correlations and discriminability
for the multidimensional solutions.
Discrimination Between Factors
For the two-factor solution, the correlation
between the constructs of LLL (consisting of
grammar and vocabulary measures) and discourse
skill (D) was very high at .91. Brown (2006) rec-
ommends factor correlations lower than .80 or .85
for good discrimination of latent constructs. Thus,
discrimination among the two constructs can be
characterized as poor, and it is not clear that a
two-dimensional representation of language ability
provided a substantive improvement over that of
the unidimensional model. We found similar
results for the factors in the three-dimensional
model representing grammar (G), vocabulary (V),
and discourse (D). All factor correlations were
very high at .90 or larger for each pair of latent
variables, suggesting that discrimination among
these constructs was weak for the three-dimen-
sional model as well.
1958 LARRC

Table 3
Grade-Specific Average Variance Extracted (AVE) by Each Factor, and Squared Factor Correlations (SFCs) for the One-, Two-, and Three-Dimen-
sional Models of Language

Two-dimensional model Three-dimensional model

One-dimensional model
AVE AVE
SFC (%) SFC (%) SFC (%) SFC (%)
Grade AVE LLL D LLL versus D G V D G versus D V versus D G versus V

PK 45.4 48.9 43.2 83.4 47.4 50.7 43.2 80.6 87.6 99.6
K 48.7 52.6 43.6 90.6 52.4 54.6 43.7 83.9 95.5 95.3
G1 46.8 53.5 43.0 72.3 61.6 41.0 43.0 67.7 90.6 > 100.0a
G2 44.6 51.6 43.7 64.5 57.9 43.2 43.8 61.6 73.6 > 100.0a
G3 43.2 55.3 29.8 61.3 56.4 62.6 29.8 65.3 48.9 80.6

Note. PK = prekindergarten; K = kindergarten; G1 = Grade 1; G2 = Grade 2; G3 = Grade 3; LLL = lower level language, D = discourse,
G = grammar, V = vocabulary..
a
Inadmissible solution.

Standardized factor loadings in all three models

of language dimensionality (not shown) were statis-
tically significant and the majority of loadings in
each CFA were > 0.70. All but one measure had stan-
dardized loadings > 0.50; the lowest standardized
loading in the PK CFAs was 0.49 in the unidimen-
sional model for the past tense probe of the TEGI.
For each factor, the factor loadings can be squared
and summed to find the AVE; when compared to the
SFCs, the AVE provides a more rigorous assessment
of discrimination among factors than simple review
of the factor correlations (Fornell & Larcker, 1981;
Hair et al., 2010; Netemeyer et al., 2003).
The top row of Table 3 contains the AVE for the
PK data across the factors for the one-, two-, and
three-dimensional models. For the unidimensional
model, the AVE for the single factor was 45.4%. In
the two-dimensional model, the LLL construct is
able to explain an average of 48.9% of the variance
in the set of its indicator G and V measures, with D
explaining an average of 43.2% of the variance in
its set of discourse skill indicators. These AVEs are
substantially lower than the variance shared
between the constructs of LLL and D, which was
83.4%, and thus discrimination between factors is
not supported (e.g., Netemeyer et al., 2003). This
finding is similar to the results for the three-factor
model. All three AVEs for the three-factor G, V,
and D model are substantially smaller than the
shared variance between these factors. The latent
variables in the two- and three-dimensional repre-
sentations of language considered here shared more
variance between them than they were able to
extract from their corresponding set of measures;
thus, the latent factors in the multidimensional
models are not reasonably distinct.
Summary: PK Results
For the three specific models we tested, results of
the chi-square comparison tests support the two-
dimensional model for language in PK. However,
we note some ambiguity in claiming that the collec-
tion of evidence is strong in support of this model.
First, fit criteria were acceptable and nearly equal
across all of the models, and the CFI in particular
was no different for these models, providing some
evidence in favor or the more parsimonious solution
(i.e., one-dimensional model; Moran et al., 2013).
Second, factor correlations are very high for the two-
dimensional model, and investigation of the AVE by
each factor relative to the squared multiple correla-
tions between factors suggests that meaningful dif-
ferences between the two dimensions are not readily
discerned. We estimated construct reliability for the
uni- and two-dimensional models using coefficient
H (Hancock & Mueller, 2001). For the unidimen-
sional model, reliability was high with H = .93. For
the two-dimensional model, reliability was still high
for the LLL factor (grammar and vocabulary skills),
H = .91, but was lower for the D factor, H = .782.
Although an argument could be made in support of
the two-dimensional model based strictly on the
scaled chi-square difference tests, it seems unlikely
that the two-dimensional model yields a stronger
substantive interpretation to the data over the parsi-
mony available through the one-dimensional model,
at least for the specific models that we have com-
pared here at PK. Thus, from a substantive perspec-
tive, the unidimensional model at PK seems
preferable to the two separate but highly correlated
factors (rLLL-D = .91, with r2 = 83.4%) obtained for
the two-dimensional model.

Kindergarten: Dimensionality of Language Ability
Model Fit Criteria
Examination of model fit criteria showed no clear
preference for one representation of language over
another (with more consistency across models than
the fit results for the PK data; see Table 2).
Nested Model Comparisons
The unidimensional model of language in
kindergarten was retained, as none of the adjusted
chi-square difference tests were statistically signifi-
cant (Table 2, adjusted Dv2).
Discrimination Between Factors
Further evidence supporting unidimensionality
of language for the kindergarten data comes from
examination of the factor correlations in the multi-
dimensional models, which are all very large (last
column of Table 2) and are above the recommenda-
tions for good discrimination of latent constructs in
a multifactor model. Standardized factor loadings
in all three models of language dimensionality (not
shown) were statistically significant and the major-
ity of loadings in each CFA were > 0.70. Three
measures had loadings below 0.50 in the unidimen-
sional and two-dimensional models, and only two
had loadings below 0.50 in the three-dimensional
model. As shown in Table 3, the AVEs for the
latent factors in the two- and three-dimensional
models are less than the large proportion of vari-
ance each factor shares with the other factor(s) in
that model (SFCs); thus, in kindergarten there is no
evidence or support for distinct constructs of LLL
versus D, or for G versus V versus D ability.
Summary: K Results
Findings suggest that the best representation of
dimensionality of language ability at kindergarten
is the unidimensional model. Construct reliability
was high, with H = .94.
G1: Dimensionality of Language Ability
Model Fit Criteria
Examination of model fit criteria suggests that
we can discount the three-factor model due to
inadmissible values for one of the factor correla-
tions, with a correlation > 1.0 between the latent
Dimensionality of Language 1959
constructs for G and V (rG-V = 1.06). This is not an
atypical occurrence in CFA and suggests that model
modifications are warranted. Given our specific ser-
ies of theoretical and confirmatory models to be
compared, we did not consider adaptations to the
three-dimensional model. The two-dimensional
model fit criteria were slightly improved over the
one-dimensional model (Table 2, third section). For
example, the chi-square value was smaller, the p-
close value was larger, the CFI was larger, and the
AIC was smaller for the two-dimensional model.
SRMR was similar for the uni- and two-dimen-
sional models. None of the models attained sug-
gested cutoffs for the RMSEA. Despite larger values
on the RMSEA for all three values, the remaining
criteria suggest acceptable model fit for the two-
dimensional model: LLL versus D ability.
Nested Model Comparisons
A statistical comparison of the uni- versus two-
dimensional models of language (Table 2) found
that the unidimensional model fit significantly
worse than the two-dimensional model,
v2(1) = 17.82, p < .01. Thus, there is support for the
two-factor solution.
Discrimination Between Factors
The correlation between the LLL and D ability
factors in the two-factor solution is quite high,
rLLL-D = .85, calling into question the distinctiveness
of these two factors. However, Brown (2006) notes
that .85 (or more conservatively, .80) serves as an
appropriate upper bound for discriminability
among factors; thus, although high, separate contri-
butions to language by these two factors are likely
to be substantively meaningful. To further under-
stand this overlap, we reviewed the discriminant
validity properties for the two-dimensional model.
Standardized factor loadings in all three models of
language dimensionality (not shown) were statisti-
cally significant and the majority of loadings in
each CFA were > 0.70. Two measures had loadings
below 0.50 in the unidimensional model, and only
one was below this criteria in the two-dimensional
and the three-dimensional models. The AVE results
(Table 3) indicate that the variance extracted for
each of the two factors in the two-dimensional
model is still lower than the variance shared
between the two factors. Although we would like
to see the AVEs be larger than the SFCs for distinct
constructs of LLL versus D, in G1 relative to K and
PK, the amount of shared variance (SFCs) in the
1960 LARRC
two-dimensional models is decreasing, and the
amount of variance extracted by each factor (AVEs)
in these models is somewhat increasing.
Summary: G1 Results
Considering the model fit criteria and the chi-
square difference test, the two-dimensional model
of language ability at G1 is preferred. Construct
reliabilities are moderate to high; for LLL, H = .92,
and for D, H = .76. There is considerable overlap
between the constructs of LLL and D skills
(r2 = 72.3%) but despite this overlap, we begin to
see an emergence of language dimensionality speci-
fic to LLL and D skills at first grade.
G2: Dimensionality of Language Ability
Model Fit Criteria
Similar to the G1 results, the three-factor model
can be discounted due to inadmissible values for
one of the factor correlations (rG-V = 1.04 between
the latent factors for G and V in the three-dimen-
sional solution). Comparing across the uni- and
two-dimensional results (Table 2), the model fit cri-
teria were better for the two-dimensional model.
The upper CI boundary on the RMSEA was slightly
beyond 0.08, but all other criteria were within
acceptable limits. Thus, based strictly on these crite-
ria, there is support for the two-factor solution.
Nested Model Comparisons
A statistical comparison of the one- versus two-
dimensional models of language (Table 2, far-right
column) yielded preference for the two-factor solu-
tion, v2(1) = 17.97, p < .01.
Discrimination Between Factors
As we saw with the G1 sample, the correlation
between the two factors of LLL and D was high,
rLLL-D = .80, but within an acceptable range for dis-
criminability between the factors. Standardized fac-
tor loadings in all three models of language
dimensionality (not shown) were statistically signif-
icant and the majority of loadings in each CFA
were > 0.70. Two measures had loadings below
0.50 in the uni- and two-dimensional models, and
only one was below this criteria in the three-dimen-
sional model. The AVE results (Table 3) indicate
that the AVEs for each of the two factors in the
two-dimensional model (51.6% and 43.7%,
respectively) were lower than the variance shared
between the two factors (64.5%). This discrepancy
was not what we would desire for distinct con-
structs of LLL versus D language skills, but there
was continuing evidence across grades that the
amount of shared variance (SFCs) in the two-di-
mensional models was gradually decreasing, and
the strength of these language skills as independent
dimensions of language (AVE) was improving.
Summary: G2 Results
Based on the model fit criteria and the chi-square
difference test, the two-dimensional model best rep-
resents language ability at G2. Construct reliabilities
were moderate to high; for LLL, H = .92, and for D,
H = .77. We note again the considerable overlap
between the constructs of LLL and D skills
(r2 = 64.5%) at second grade, but continue to detect
evidence for the emergence of differentiation
between these two constructs.
G3: Dimensionality of Language Ability
Model Fit Criteria
The final section of Table 2 contains model fit
statistics for the three confirmatory factor models
for the G3 sample. The three-dimensional solution
is interpretable here, with admissible parameter
estimates for the factor correlations. All of the
model fit criteria were best for the three-dimen-
sional solution; we note that the upper bound on
the RMSEA confidence interval was, however, out-
side the recommended 0.08 cutoff.
Nested Model Comparisons
The two-dimensional and three-dimensional
models yielded better fit compared to the one-
dimensional model, v2(1) = 7.51, p < .01 and
v2(3) = 12.87, p < .01, respectively (see Table 2), but
there was also a statistical difference between the
two- and three-dimensional models, v2(2) = 6.19,
p < .05. Based on these results, the representation of
language in the three-dimensional model that sepa-
rates G, V, and D language skills fit the data signifi-
cantly better than either lower order solution.
Discrimination Between Factors
Factor correlations, however, were high in both
the two- and three-dimensional solutions, suggest-
ing that while there may be a statistical preference

for a three-dimensional model that isolates the lan-
guage skills of G, V, and D, there were also rela-
tively strong correlations between these constructs.
In particular, the strongest correlation in this set
was between the constructs of G and V in the
three-dimensional model (rG-V = .90), formerly part
of the same construct in the two-dimensional mod-
els supported in G1 and G2. Standardized factor
loadings in all three models of language dimension-
ality (not shown) were statistically significant and
the majority of loadings in each CFA were > 0.70.
Two measures had loadings below 0.50 in the uni-
dimensional model, and only one was below this
criteria in the two- and three-dimensional models.
The AVE and SFC results provided in Table 3 indi-
cate that the AVE for the first two of the three fac-
tors in the three-dimensional model (56.4%, 62.6%,
and 29.8%, respectively) were somewhat close to
but not exceeding the variance shared between the
three factors; there was also a very large 80.6% esti-
mate of variance shared between the latent vari-
ables of G and V in the three-dimensional model.
Thus, despite statistical preference for the three-
dimensional solution, overlap among at least two
factors is large. However, in the three-factor compo-
sition these three separate components of language
skill—namely, G, V, and D—seem to be emerging
as independent dimensions of language in older
children. We note, however, that the average vari-
ance explained by D in the three-dimensional
model is lower here than was obtained in the previ-
ous grades, suggesting that one or more of the indi-
cators of D may not adequately represent this
construct well in G3.
Summary: G3 Results
Based on the model fit criteria and the chi-square
difference test, the three-dimensional model was
preferred for third grade, with a cautionary note on
the overlap particularly between the G and V fac-
tors in the three-dimensional model. Construct reli-
abilities were moderate; for G, H = .85; for V,
H = .87; and for D, H = .65. In G3, differentiation of
G and V as separate skills from D ability appeared
to occur, although these skills were correlated.
Discussion
This examination of the dimensionality of language
ability in 4- to 8-year-old children significantly
extends our understanding of language develop-
ment during this critical period of early childhood
Dimensionality of Language 1961
in several important ways. First, we found that
when discourse skills were added to the study of
language skills, there was some ambiguity in the
identification of dimensionality among the youngest
children in our study (children in prekindergarten).
By kindergarten, our data showed that the con-
structs we were investigating, namely, grammar,
vocabulary, and discourse, formed a single con-
struct. Second, we demonstrate a potentially emer-
gent dimensional structure of language, resulting in
the separation of vocabulary and grammar, which
hang together, and discourse at first and second
grades, followed by differentiation of vocabulary,
grammar, and discourse by 8 years of age.
Concerning the first major finding, this work
built upon an earlier examination of the dimension-
ality of language in vocabulary and grammar (see
Tomblin & Zhang, 2006). That work indicated that
language ability is initially unidimensional but
becomes increasingly multidimensional as children
move across the primary grades and into middle
school. Complementing those initial findings, the
present work looked intensively at the dimensional-
ity of language ability for children who were 4–
8 years of age, corresponding to the time period in
which children typically enter formal schooling (at
prekindergarten) and transition through periods of
learning to read to reading to learn (third grade).
Our results help us to understand more explicitly
the nature of children’s language ability during
these formative years of schooling and reading
development. For the youngest children in our sam-
ple, language dimensionality was found to be
somewhat unclear. Specifically, for children in
prekindergarten, a two-dimensional representation
of language emerged, whereas for kindergarten
children, language was observed as a single dimen-
sion of development. Although our results can be
considered more robust than previous work in this
area due to the latent approach utilized in analyses,
we acknowledge these findings are tentative given
the small size of our kindergarten sample. Consid-
ering the instability of language observed for the
youngest children in our study, future research,
particularly work that is longitudinal in nature, is
warranted to disentangle our mixed findings
regarding the possibility that discourse-level skills
may in fact be separable from other language con-
structs at very early ages. For the older children,
language ability was best represented as two- or
even three-related dimensions. In fact, by third
grade, we see that vocabulary, grammar, and dis-
course appear to be emerging as distinct skills, and
that discourse language skills, such as inference
1962 LARRC
and comprehension monitoring, are a distinguish-
able dimension of language.
These findings of a unitary dimension for vocab-
ulary and grammar until third grade are not dis-
crepant with those of the earlier work. Although
Tomblin and Zhang (2006) preferred the two-factor
solution only for children in eighth grade, the CFI
values for their two-factor model for the earlier
grades all indicated a good fit. Together, these two
studies with different samples and test instruments
similarly show the emergence of separate constructs
of vocabulary and grammar. Thus, although these
aspects of language are often grouped together as
LLL skills, the two studies indicate that this level of
description may be relevant only for younger chil-
dren. Furthermore, this finding fits with the theoret-
ical conceptualizations of vocabulary and grammar
as lower level or foundational skills that support
the reading comprehension later (Lepola et al.,
2012).
The emergent dimensionality seen among young
children in their development of language makes it
clear that language is a complex construct. Our data
support the viewpoint that there are dimensions
within this construct that are, to some extent, sepa-
rable. However, why are these dimensions not
wholly apparent until later in children’s develop-
ment? One explanation for our finding of emergent
dimensionality is that a hierarchy of language
exists, from words to sentences to discourse (Tom-
blin & Zhang, 2006). As the dimensions lower
down in the hierarchy become consolidated and
automatized in use, as occurs early in development,
cognitive resources are freed up to support more
complex processing, as would be required for com-
prehension monitoring and inference. This may
arise because a critical base of vocabulary is needed
for grammar and, similarly, a critical base of gram-
mar is needed to understand discourse. As noted in
our Introduction, different phases of grammar
development have been linked to lexical knowledge
during the first 2 years of life (Marchman & Bates,
1994).
We do not believe that this is an adequate expla-
nation because there is evidence for the discourse
language skills that we discuss (comprehension
monitoring, knowledge of narrative, and inference)
in the first 3–4 years of life (Skarakis-Doyle &
Dempsey, 2008; Tomasello et al., 2007). An alterna-
tive explanation as to why this emergent dimen-
sionality may occur at a particular time point, is
that language abilities are likely affected by chil-
dren’s literacy experiences within schooling and the
increasing demand to apply linguistic resources to
higher level interactions with texts (orally or read)
as a major part of reading instruction. To this point,
consider that during the course of early language
development, the written texts that children experi-
ence show gradual increases in the complexity of
the morphological structure of words, the syntactic
sophistication of sentences, and the number of
propositions that have to be integrated across the
text as a whole. Children’s interactions with these
increasingly complex written texts during the first
few years of literacy instruction may enhance the
structure of language dimensions that we find.
Consistent with this argument is that the increasing
challenge of texts in the early school years provides
one explanation for the identification of a group of
poor readers whose problems only become appar-
ent in fourth grade as these children struggle to
comprehend what they read (Catts, Compton,
Tomblin, & Bridges, 2012). Similarly, it has been
shown that interactions with literacy drive phone-
mic awareness (Castles & Coltheart, 2004); thus,
schooling impacts on the structure and refinement
of oral, as well as written language. This noted, the
results here do not imply that discourse language
skills, such as comprehension monitoring and infer-
ence, are not observed among young children as
we have already discussed; rather, it appears to be
that case that around first grade these begin to
clearly represent a separable dimension of language
ability from lower level skills.
The second major finding of this work concerns
the clear emergence of discourse-level language
ability as a specific dimension of language ability
among the older children in this sample. At first
grade, we found two distinct dimensions of lan-
guage ability—one representing the lower level
skills of vocabulary and grammar and the other
representing discourse-level skills, and this distinc-
tiveness held through third grade, albeit with sepa-
ration of lower level skills (vocabulary and
grammar) observed for the third graders. This is a
unique finding because no previous studies of lan-
guage dimensionality have included assessment of
skills essential to discourse.
Limitations of this work warrant note, and we
highlight three that are most salient. The first con-
cerns the measures used to represent discourse skill.
Several of these measures have never before been
used with children as young as those in this study,
and thus further examination of their validity with
the younger groups is needed. In addition, reliabil-
ity measures for 3 of the 20 measures of discourse
used across all grades were found to be below com-
monly accepted cutoffs. Specifically, the alpha for

the experimental discourse measure, detecting
inconsistencies, was found to be particularly low
(Cronbach’s alpha = .54 in third grade). However,
in each of the models where discourse skill formed
its own dimension, the overall construct reliabilities
for the discourse factor as measured through coeffi-
cient H (Hancock & Mueller, 2001) were acceptable,
at .76 and .77 for first and second grades, respec-
tively, although somewhat lower for third grade at
.65. While additional measurement work should be
undertaken to improve the DI measure, construct
reliability is not diminished by the addition of a
weak indicator to a measurement model that
includes a collection of strong indicators (Hancock
& Mueller, 2001). Nonetheless, the measures reflect-
ing discourse skill overall could be improved. The
second limitation concerns the cross-sectional nat-
ure of the data. Prior work examining the dimen-
sionality of language ability has used longitudinal
methods rather than the cross-sectional approach
used here (Tomblin & Zhang, 2006). A limitation of
the cross-sectional design is that it essentially pro-
vides only a “snapshot” of a phenomenon at a
given point in time. However, given that some of
the measures across grades varied, examining the
models cross sectionally is a reasonable approach to
understanding dimensionality at each grade. While
we may infer that differences seen in the dimen-
sionality of language ability would be borne out
with longitudinal assessment, this requires addi-
tional empirical investigation.
A third limitation is that different theoretical
explanations could be considered. It could be
argued that the measures of discourse are simply
“more difficult,” perhaps because they place a
greater demand on cognitive resources, such as
working memory, and that if more challenging
vocabulary or grammar assessments were included,
the three-dimensional split would not arise. There
is some evidence for this viewpoint. A recent study
of children with SLI found that performance on
subtests of standardized oral language assessments
did not neatly load onto separate factors of seman-
tics and syntax or comprehension versus produc-
tion language; instead a factor structure that
reflected the different language processing require-
ments of the tasks was apparent (Hoffman, Loeb,
Brandel, & Gillam, 2011).
While this work is largely meant to be theoretical
and descriptive in nature, designed to advance
understanding of the nature of language ability
among young children, the results may have impli-
cations to practice. First, from an assessment
perspective, this study does not support the reliance
Dimensionality of Language 1963
on subtests designed to assess specific dimensions
of LLL ability (i.e., vocabulary and grammar) for
prekindergarteners through second graders, as these
do not appear to reflect specific, distinguishable
abilities. Here, we are echoing a point made by
Tomblin and Zhang (2006) for older children.
Rather, for children at these ages, their vocabulary
and grammar reflect a single underlying trait, and it
seems best to rely on omnibus assessments of lan-
guage skill to explore individual differences among
children and, potentially, to identify children who
have underdeveloped language skills warranting
attention. Second, with respect to instruction, the
study supports the notion that for first- through
third-grade children, lower and discourse-level lan-
guage skills reflect distinguishable underlying traits.
While there is increasing awareness of the impor-
tance of discourse-level language skills to reading
development, it is not clear that educators often seek
to identify children who may have limitations in
this dimension of language ability. Research show-
ing that discourse skills appear underdeveloped in
children who have comprehension-specific reading
deficits in the later primary grades (e.g., Adlof,
Catts, & Little, 2006) argues for the importance for
monitoring these skills early in development.
References
Adlof, S., Catts, H., & Little, T. (2006). Should the sim-
ple view of reading include a fluency component?
Reading and Writing, 19, 933–958. doi:10.1007/s11145-
006-9024-z
American Psychiatric Association. (2012). Diagnostic and
statistical manual of mental disorders. Arlington, VA:
Author.
Bates, E., Dale, P., & Thal, D. (1995). Individual differ-
ences and their implications for theories of language
development. In P. Fletcher & B. MacWhinney (Eds.),
Handbook of child language (pp. 96–151). Oxford, UK:
Blackwell.
Berninger, V. W., & Abbott, R. D. (2010). Listening com-
prehension, oral expression, reading comprehension,
and written expression: Related yet unique language
systems in Grades 1, 3, 5, and 7. Journal of Educational
Psychology, 102, 635–651. doi:10.1037/a0019319
Bishop, D. V. M. (1997). Uncommon understanding: Devel-
opment and disorders of language comprehension in chil-
dren. Hove, UK: Psychology Press.
Bishop, D. V. (2003). Test for reception of grammar (2nd
ed.). Minneapolis, MN: Pearson.
Brown, T. A. (2006). Confirmatory factor analysis for applied
research. New York, NY: Guilford.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of
assesssing model fit. In K. A. Bollen & J. S. Long (Eds.),
1964 LARRC
Testing structural equation models (pp. 136–162). New-
bury Park, CA: Sage.
Byrne, B. M. (2012). Structural equation modeling with
Mplus. New York, NY: Routledge.
Cain, K., & Oakhill, J. (2006). Profiles of children with
specific reading comprehension difficulties. British Jour-
nal of Educational Psychology, 76, 683–696. doi:10.1348/
000709905X67610
Cain, K., Oakhill, J., & Bryant, P. (2004). Children’s read-
ing comprehension ability: Concurrent prediction by
working memory, verbal ability, and component skills.
Journal of Educational Psychology, 96, 31–42. doi:10.1037/
0022-0663.96.1.31
Castles, A., & Coltheart, M. (2004). Is there a causal link
from phonological awareness to success in learning to
read? Cognition, 91, 77–111.
Catts, H. W., Compton, D., Tomblin, J. B., & Bridges, M.
S. (2012). Prevalence and nature of late-emerging poor
readers. Journal of Educational Psychology, 104, 166–181.
doi:10.1037/a0025323
Catts, H. W., Fey, M. E., Zhang, X., & Tomblin, J. B. (1999).
Language basis of reading and reading disabilities: Evi-
dence from a longitudinal investigation. Scientific Studies
of Reading, 3, 331–361. doi:10.1207/s1532799xssr0304_2
Catts, H. W., Hogan, T. P., & Fey, M. E. (2003).
Subgrouping poor readers on the basis of individual
differences in reading-related abilities. Journal of
Learning Disabilities, 36, 151–164. doi:10.1177/002221940
303600208
Dickinson, D., Golinkoff, R., & Hirsh-Pasek, K. (2010).
Speaking out for language: Why language is central to
reading development. Educational Researcher, 39, 305–
310. doi:10.3102/0013189X10370204
Dunn, L. M., & Dunn, D. M. (2007). Peabody Picture Vocab-
ulary Test (4th ed.). Minneapolis, MN: Pearson.
Elsabbagh, M., & Karmiloff-Smith, A. (2006). Modularity
of mind and language. In K. Brown (Eds.), Encyclopae-
dia of Language and Linguistics (pp. 218–224). Amster-
dam, Netherlands: Elsevier Science.
Fornell, C., & Larcker, D. F. (1981). Structural equation
models with unobservable variables and measurement
error: Algebra and statistics. Journal of Marketing
Research, 382–388. doi:10.2307/3150980
Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E.
(2010). Multivariate data analysis: A global perspective (7th
ed.). Upper Saddle River, NJ: Pearson.
Hancock, G. R., & Mueller, R. O. (2001). Rethinking
construct reliability within latent variable systems. In S.
Du Toit, R. Cudeck & D. Sorbom (Eds.), Structural
Equation Modeling: Present and Future (pp.195–216). Chi-
cago, IL: Scientific Software International.
Hancock, G. R., & Mueller, R. O. (2010). Structural equa-
tion modeling. In G. R. Hancock & R. O. Mueller
(Eds.), The reviewer’s guide to quantitative methods in the
social sciences (pp. 371–384). New York, NY: Routledge.
Hancock, G. R., & Mueller, R. O. (Eds.). (2013). Structural
equation modeling: A second course (2nd ed.). Charlotte,
SC: IAP.
Hoffman, L. M., Loeb, D. F., Brandel, J., & Gillam, R. B.
(2011). Concurrent and construct validity of oral lan-
guage measures with school-age children with specific
language impairment. Journal of Speech, Language, and
Hearing Research, 54, 1597–1608. doi:10.1044/1092-4388
(2011/10-0213)
Hogan, T. P., Cain, K., & Bridges, M. S. (2012). Young
children’s oral language abilities and later reading com-
prehension. In T. Shanahan & C. J. Lonigan (Eds.),
Early Childhood Literacy: The National Early Literacy Panel
and Beyond (pp. 217–232). Baltimore, MD:: Brookes.
Hu, L. T., & Bentler, P. M. (1995). Evaluating model fit.
In R. H. Hoyle (Ed.), Structural equation modeling: Con-
cepts, issues and applications (pp. 76–99). Thousand Oaks,
CA: Sage.
Hu, L.-T., & Bentler, P. M. (1999). Cutoff criteria for fit
indexes in covariance structure analysis: Conventional
criteria versus new alternatives. Structural Equation
Modeling, 6, 1–55.
Kim, Y. S. (2014). Language and cognitive predictors of
text comprehension: Evidence from multivariate analy-
sis. Child Development. Advance online publication.
doi:10.1111/cdev.12293
Kline, R. (2013). Exploratory and confirmatory factor anal-
ysis. In Y. Petscher, C. Schatschneider, & D. Compton
(Eds.), Applied quantitative analysis in education and the
social sciences (pp. 171–207). New York, NY: Routledge.
Lepola, J., Lynch, J., Laakkonen, E., Silven, M., & Niemi,
P. (2012). The role of inference making and other lan-
guage skills in the development of narrative listening
comprehension in 4–6-year-old children. Reading
Research Quarterly, 47, 259–282. doi:10.1002/rrq.020
Lomax, R. (2013). Introduction to structural equation
modeling. In Y. Petscher, C. Schatschneider, & D.
Compton (Eds.), Applied quantitative analysis in education
and the social sciences (pp. 245–264). New York, NY:
Routledge.
MacCallum, R. C., & Austin, J. T. (2000). Applications of
structural equation modeling in psychological research.
Annual Review of Psychology, 51, 201–226. doi:10.1146/
annurev.psych.51.1.201
Marchman, V. A., & Bates, E. (1994). Continuity in lexical
and morphological development: A test of the critical
mass hypothesis. Journal of Child Language, 21, 339–366.
doi:10.1017/S0305000900009302
McCoach, D. B., Black, A. C., & O’Connell, A. A. (2007).
Errors of inference in structural equation modeling.
Psychology in the Schools, 44, 461–470. doi:10.1002/
pits.20238
Mills, D. L., Coffey-Corina, S., & Neville, H. J. (1997).
Language comprehension and cerebral specialization
from 13 to 20 months. Developmental Neuropsychology,
13, 397–445. doi:10.1080/87565649709540685
Moran, A. J. S., Marsh, H. W., & Nagengast, B. (2013).
Exploratory structural equation modeling. In G. R.
Hancock & R. O. Mueller (Eds.), Structural equation
modeling: A second course (2nd ed., pp. 1–10). Charlotte,
SC: IAP.

Mueller, R. O., & Hancock, G. R. (2010). Structural equa-
tion modeling. In G. R. Hancock & R. O. Mueller
(Eds.), The reviewer’s guide to quantitative methods in the
social sciences (pp. 371–383). New York, NY: Routledge.
Muthen, L. K., & Muthen, B. O. (1993–2012). MPlus user’s
guide (7th ed.). Los Angeles, CA: Author.
Nation, K., Cocksey, J., Taylor, J. S. H., & Bishop, D. V.
M. (2010). A longitudinal investigation of early reading
and language skills in children with poor reading com-
prehension. Journal of Child Psychology and Psychiatry,
51, 1031–1039. doi:10.1111/j.1469-7610.2010.02254
Netemeyer, R. G., Bearden, W. O., & Sharma, S. (2003).
Scaling procedures: Issues and applications. Thousand
Oaks, CA: Sage.
Oakhill, J. V., & Cain, K. (2012). The precursors of read-
ing ability in young readers: Evidence from a four-year
longitudinal study. Scientific Studies of Reading, 16,
91–121. doi:10.1080/10888438.2010.529219
Pinker, S. (1998). Words and rules. Lingua, 106, 219–242.
Plunkett, K., & Marchman, V. (1993). From rote learning
to system building: Acquiring verb morphology in chil-
dren and connectionist nets. Cognition, 48, 21–69.
doi:10.1111/j.1469-7610.2010.02254.x
Rice, M. L., & Wexler, K. (2001). Rice/Wexler Test of Early
Grammatical Impairment. San Antonio, TX: Psychological
Corporation.
Satorra, A., & Bentler, P. M. (1994). Corrections to test
statistics and standard errors in covariance structure
analysis. In A. v. E. C. C. Clogg (Ed.), Latent variables
analysis: Applications for developmental research (pp. 399–
419). Thousand Oaks, CA: Sage.
Schumacker, R. E., & Lomax, R. G. (2004). A beginner’s
guide to structural equation modeling. New York, NY:
Psychology Press.
Schwartz, R. G. (2009). Handbook of child language disor-
ders. New York, NY: Psychology Press.
Semel, E., Wiig, E. H., & Secord, W. A. (2003). Clinical
Evaluation of Language Fundamentals (4th ed.). Blooming-
ton, MN: Pearson.
Dimensionality of Language 1965
Silva, M. T., & Cain, K. (2015). The relations between
lower- and higher-level oral language skills and their
role in prediction of early reading comprehension. Jour-
nal of Educational Psychology, 107, 321–331. doi:10.1037/
a0037769
Skarakis-Doyle, E., & Dempsey, L. (2008). The detection
and monitoring of comprehension errors by preschool
children with and without language impairment. Jour-
nal of Speech, Language, and Hearing Research, 51, 1227–
1243. doi:10.1044/1092-4388(2008/07-0136)
Stein, N. L., & Glenn, C. G. (1982). Children’s concept
of time: The development of a story schema. In W.
J.. Friedman & H. Belin (Eds.), The Developmental Psy-
chology of Time (pp. 255–282). Waltham, MA: Aca-
demic Press. doi:10.1146/annurev.clinpsy.1.102803.
144239
Tomarken, A. J., & Waller, N. G. (2003). Potential pro-
blems with "well fitting" models. Journal of Abnormal
Psychology, 112, 578–598.
Tomarken, A. J., & Waller, N. G. (2005). Structural equa-
tion modeling: Strengths, limitations, and misconcep-
tions. Annual Review of Clinical Psychology, 1, 31–65.
Tomasello, M., Carpenter, M., & Liszkowski, U. (2007). A
new look at infant pointing. Child Development, 78, 705–
722. doi:10.1111/j.1467-8624.2007.01025.x
Tomblin, J. B., & Zhang, X. (2006). The dimensionality
of language ability in school-age children. Journal
of Speech, Language, and Hearing Research, 49, 1193–
1208.
Wagner, R. K. (n.d.). Morphological Derivation Task. Talla-
hassee: Florida State University.
Wechsler, D., & Kort, W. (2005). WISC–IIInl: Wechsler
Intelligence Scale for Children. San Diego, CA: Harcourt
Assessment.
Wiig, E. H., Secord, W. A., & Semel, E. (2004). Clinical
Evaluation of Language Fundamentals Preschool (2nd ed.).
San Antonio, TX: Harcourt Assessment.
Williams, K. T. (2007). Expressive Vocabulary Test (2nd
ed.). Bloomington, MN: Pearson.